Qualcomm has landed a major hyperscale customer for custom data center silicon, marking the company’s most aggressive push into server infrastructure since it abandoned that market entirely in 2018.
Shipments of the new custom chips are expected to begin in December 2026.
From mobile giant to data center contender
Qualcomm launched its Centriq server processors in 2017, positioning ARM-based chips as a power-efficient alternative to Intel’s dominance. By 2018, the entire effort was shuttered.
The AI inference play
Rather than trying to compete head-on with NVIDIA’s training GPUs, Qualcomm is targeting a specific slice of the AI compute stack: inference.
The company is developing custom ASIC-based AI accelerators designed to handle inference workloads at lower power consumption than GPU-heavy alternatives.
Qualcomm signaled this strategic direction in August 2025, and the signing of an unnamed hyperscaler validates that the pitch is landing with at least one major customer.
A crowded and complicated competitive field
NVIDIA remains the undisputed leader, with its GPU architecture serving as the default platform for both training and inference. AMD has gained ground with its MI-series accelerators. Intel continues to push its Gaudi accelerators.
Amazon has built its own Trainium and Inferentia chips for AWS. Google has its TPUs. Microsoft is developing its own Maia accelerators.
Reports also indicate Qualcomm is exploring AI inference hardware deployment in new geographic regions, particularly Latin America.
The December 2026 shipment timeline means the company needs to execute on custom silicon development. One hyperscale customer is a proof point, not a business. Qualcomm will need to demonstrate that its inference architecture can attract multiple large buyers to justify the R&D investment.
Be the first to comment